[Xenomai-core] Re: [PATCH] kgdb/x86 over I-pipe

2006-06-15 Thread Jan Kiszka
Philippe Gerum wrote:
 Jan Kiszka wrote:
 Philippe Gerum wrote:
 While we are at it,
 #define current ipipe_safe_current() /* ? */


 Nope, there is the need for some special changes.

 
 If you refer to the cache flushing issue, then it would be better to
 actually check for foreign stacks explicitely, so that you could
 substitute current globally:
 
 -if (CACHE_FLUSH_IS_SAFE  current-mm 
 -addr  TASK_SIZE)
 +if (CACHE_FLUSH_IS_SAFE  !testbit(IPIPE_NOSTACK_FLAG, 
 + ipipe_percpu_domain[cpuid]-cpudata[cpuid].status) 
 +current-mm  addr  TASK_SIZE)
 

Right - on first sight. I tried to redefine current, but the
ipipe_safe_current macro requires that symbol itself, ugh. Turning
ipipe_safe_current into a static inline doesn't work due to circular
dependencies on linux/sched.h.

So I guess it's best to keep it as it is (though kgdb-ipipe.patch would
have become really cute).

Jan



signature.asc
Description: OpenPGP digital signature
___
Xenomai-core mailing list
Xenomai-core@gna.org
https://mail.gna.org/listinfo/xenomai-core


[Xenomai-core] Re: [PATCH] kgdb/x86 over I-pipe

2006-06-15 Thread Philippe Gerum

Jan Kiszka wrote:

Philippe Gerum wrote:


Jan Kiszka wrote:


Philippe Gerum wrote:


While we are at it,
#define current ipipe_safe_current() /* ? */



Nope, there is the need for some special changes.



If you refer to the cache flushing issue, then it would be better to
actually check for foreign stacks explicitely, so that you could
substitute current globally:

-if (CACHE_FLUSH_IS_SAFE  current-mm 
-addr  TASK_SIZE)
+if (CACHE_FLUSH_IS_SAFE  !testbit(IPIPE_NOSTACK_FLAG, 
+ ipipe_percpu_domain[cpuid]-cpudata[cpuid].status) 

+current-mm  addr  TASK_SIZE)




Right - on first sight. I tried to redefine current, but the
ipipe_safe_current macro requires that symbol itself, ugh. Turning
ipipe_safe_current into a static inline doesn't work due to circular
dependencies on linux/sched.h.

So I guess it's best to keep it as it is (though kgdb-ipipe.patch would
have become really cute).



Ack.

--

Philippe.

___
Xenomai-core mailing list
Xenomai-core@gna.org
https://mail.gna.org/listinfo/xenomai-core


[Xenomai-core] Re: [PATCH] kgdb/x86 over I-pipe

2006-06-12 Thread Jan Kiszka
Philippe Gerum wrote:
 Jan Kiszka wrote:
 Philippe Gerum wrote:

 Jan Kiszka wrote:

 Philippe Gerum wrote:


 Hi Jan,

 Based on your previous work, here is a set of patches coupling KGDB
 and
 the I-pipe. Basically, I've attempted to shrink the extra patches
 needed
 against the original KGDB + I-pipe ones to the bare minimum. This has
 been obtained by having the I-pipe provide ipipe_current_safe(), and
 drastically reduce the amount of fiddling with smp_processor_id().

 The key difference with the former implementation is that a domain
 (e.g.
 Xenomai) is now expected to tell the I-pipe when it's switching to a
 non-Linux stack, and the I-pipe makes good use of this information to
 return the proper current value when asked to through
 ipipe_safe_current() from the KGDB code. The issue of swapping


 This looks nice.



 smp_processor_id() with ipipe_processor_id() has been addressed the
 hard
 way: smp_processor_id() is simply defined as ipipe_processor_id() when
 CONFIG_IPIPE and CONFIG_KGDB are both enabled in include/linux/smp.h.
 This approach was actually used during the old Adeos times when
 pipeline
 domain had their own separate stack. I take for granted that the CPU
 penalty taken in doing this is perfectly acceptable, since well, we
 are
 debugging after all.


 There is only one drawback: we will not be able to debug
 smp_processor_id-related bugs in ipipe/Xenomai anymore...


 Good point. Here is an updated patch.



 Nope that's not the point I meant. I was referring to bugs like the
 missing smp_processor_id patch in asm-i386/mmu_context.h. Your way makes
 such problems disappear once you switch on the debugger. Remember that
 we spotted the mmu_context issue via kgdb.

 
 Got it now. Indeed, making such kind of bugs disappear would be a very
 undesirable side-effect of enabling KGDB.
 
 The attached kgdb-ipipe patch (which also updates to kgdb-CVS head)
 addresses maintenance and smp-safeness by redefining the involved
 services only per source file, not kernel-wide. What's do you think
 about this approach?

 
 Mixed feelings. On one hand, it locally fixes the issue without masking
 any bugs and that's good. On the other hand, I've been bitten using this
 kind of tricks in earlier Adeos releases for patching printk() the same
 way, substituting it with spin_lock_irqsave_hw. At some point,
 spin_lock_irqsave became local_irq_save + spin_lock in the vanilla code,
 making the substitution of the former pointless, and opening a
 preemption hole in the code. This said, no spinlock macros are involved
 in the substitution you are suggesting, so there is no such risk, and in
 any case, the risk is confined to one file and not spread all over the
 place like with the raw_smp_processor_id() substitution. IOW, let's go
 your way.

I agree that it's kind of fragile, but we need to check on kgdb updates
anyway. I hope for kgdb becoming mainline one day. Then the code should
stabilise (if it hasn't already), and we can track it via ipipe directly
- without tricks. So far, this should help to keep efforts lower for us.

 
 While we are at it,
 #define current ipipe_safe_current() /* ? */

Nope, there is the need for some special changes.

 
 I'm also posting two fixes for ipipe itself to 1) make kgdb work over
 SMP and 2) fix a compiler warning.

 This version works fine for me, at least in qemu. I tried to check SMP
 as well, but qemu obviously lacks support for NMIs raised via IPI, and
 this is used by kgdb to sync all CPUs on debugging events (vanilla kgdb
 suffers too).

 

Jan



signature.asc
Description: OpenPGP digital signature
___
Xenomai-core mailing list
Xenomai-core@gna.org
https://mail.gna.org/listinfo/xenomai-core


[Xenomai-core] Re: [PATCH] kgdb/x86 over I-pipe

2006-06-12 Thread Philippe Gerum

Jan Kiszka wrote:

Philippe Gerum wrote:


Jan Kiszka wrote:


Philippe Gerum wrote:



Jan Kiszka wrote:



Philippe Gerum wrote:




Hi Jan,

Based on your previous work, here is a set of patches coupling KGDB
and
the I-pipe. Basically, I've attempted to shrink the extra patches
needed
against the original KGDB + I-pipe ones to the bare minimum. This has
been obtained by having the I-pipe provide ipipe_current_safe(), and
drastically reduce the amount of fiddling with smp_processor_id().

The key difference with the former implementation is that a domain
(e.g.
Xenomai) is now expected to tell the I-pipe when it's switching to a
non-Linux stack, and the I-pipe makes good use of this information to
return the proper current value when asked to through
ipipe_safe_current() from the KGDB code. The issue of swapping



This looks nice.





smp_processor_id() with ipipe_processor_id() has been addressed the
hard
way: smp_processor_id() is simply defined as ipipe_processor_id() when
CONFIG_IPIPE and CONFIG_KGDB are both enabled in include/linux/smp.h.
This approach was actually used during the old Adeos times when
pipeline
domain had their own separate stack. I take for granted that the CPU
penalty taken in doing this is perfectly acceptable, since well, we
are
debugging after all.



There is only one drawback: we will not be able to debug
smp_processor_id-related bugs in ipipe/Xenomai anymore...



Good point. Here is an updated patch.




Nope that's not the point I meant. I was referring to bugs like the
missing smp_processor_id patch in asm-i386/mmu_context.h. Your way makes
such problems disappear once you switch on the debugger. Remember that
we spotted the mmu_context issue via kgdb.



Got it now. Indeed, making such kind of bugs disappear would be a very
undesirable side-effect of enabling KGDB.



The attached kgdb-ipipe patch (which also updates to kgdb-CVS head)
addresses maintenance and smp-safeness by redefining the involved
services only per source file, not kernel-wide. What's do you think
about this approach?



Mixed feelings. On one hand, it locally fixes the issue without masking
any bugs and that's good. On the other hand, I've been bitten using this
kind of tricks in earlier Adeos releases for patching printk() the same
way, substituting it with spin_lock_irqsave_hw. At some point,
spin_lock_irqsave became local_irq_save + spin_lock in the vanilla code,
making the substitution of the former pointless, and opening a
preemption hole in the code. This said, no spinlock macros are involved
in the substitution you are suggesting, so there is no such risk, and in
any case, the risk is confined to one file and not spread all over the
place like with the raw_smp_processor_id() substitution. IOW, let's go
your way.



I agree that it's kind of fragile, but we need to check on kgdb updates
anyway. I hope for kgdb becoming mainline one day. Then the code should
stabilise (if it hasn't already), and we can track it via ipipe directly
- without tricks. So far, this should help to keep efforts lower for us.



While we are at it,
#define current ipipe_safe_current() /* ? */



Nope, there is the need for some special changes.



If you refer to the cache flushing issue, then it would be better to 
actually check for foreign stacks explicitely, so that you could 
substitute current globally:


-   if (CACHE_FLUSH_IS_SAFE  current-mm 
-   addr  TASK_SIZE)
+	if (CACHE_FLUSH_IS_SAFE  !testbit(IPIPE_NOSTACK_FLAG,  		+	 
ipipe_percpu_domain[cpuid]-cpudata[cpuid].status) 

+   current-mm  addr  TASK_SIZE)

--

Philippe.

___
Xenomai-core mailing list
Xenomai-core@gna.org
https://mail.gna.org/listinfo/xenomai-core


[Xenomai-core] Re: [PATCH] kgdb/x86 over I-pipe

2006-06-10 Thread Jan Kiszka
Philippe Gerum wrote:
 Jan Kiszka wrote:
 Philippe Gerum wrote:

 Hi Jan,

 Based on your previous work, here is a set of patches coupling KGDB and
 the I-pipe. Basically, I've attempted to shrink the extra patches needed
 against the original KGDB + I-pipe ones to the bare minimum. This has
 been obtained by having the I-pipe provide ipipe_current_safe(), and
 drastically reduce the amount of fiddling with smp_processor_id().

 The key difference with the former implementation is that a domain (e.g.
 Xenomai) is now expected to tell the I-pipe when it's switching to a
 non-Linux stack, and the I-pipe makes good use of this information to
 return the proper current value when asked to through
 ipipe_safe_current() from the KGDB code. The issue of swapping


 This looks nice.


 smp_processor_id() with ipipe_processor_id() has been addressed the hard
 way: smp_processor_id() is simply defined as ipipe_processor_id() when
 CONFIG_IPIPE and CONFIG_KGDB are both enabled in include/linux/smp.h.
 This approach was actually used during the old Adeos times when pipeline
 domain had their own separate stack. I take for granted that the CPU
 penalty taken in doing this is perfectly acceptable, since well, we are
 debugging after all.


 There is only one drawback: we will not be able to debug
 smp_processor_id-related bugs in ipipe/Xenomai anymore...

 
 Good point. Here is an updated patch.
 

Nope that's not the point I meant. I was referring to bugs like the
missing smp_processor_id patch in asm-i386/mmu_context.h. Your way makes
such problems disappear once you switch on the debugger. Remember that
we spotted the mmu_context issue via kgdb.

The attached kgdb-ipipe patch (which also updates to kgdb-CVS head)
addresses maintenance and smp-safeness by redefining the involved
services only per source file, not kernel-wide. What's do you think
about this approach?

I'm also posting two fixes for ipipe itself to 1) make kgdb work over
SMP and 2) fix a compiler warning.

This version works fine for me, at least in qemu. I tried to check SMP
as well, but qemu obviously lacks support for NMIs raised via IPI, and
this is used by kgdb to sync all CPUs on debugging events (vanilla kgdb
suffers too).

Jan
Index: linux/kernel/kgdb.c
===
--- linux.orig/kernel/kgdb.c
+++ linux/kernel/kgdb.c
@@ -49,6 +49,15 @@
 #include linux/console.h
 #include asm/byteorder.h
 
+#if defined(CONFIG_KGDB)  defined(CONFIG_IPIPE)
+#undef smp_processor_id
+#undef local_irq_save
+#undef local_irq_restore
+#define smp_processor_idipipe_processor_id
+#define local_irq_save  local_irq_save_hw
+#define local_irq_restore   local_irq_restore_hw
+#endif
+
 extern int pid_max;
 extern int pidhash_init_done;
 
@@ -743,7 +752,7 @@ static void kgdb_wait(struct pt_regs *re
 	local_irq_save(flags);
 	processor = smp_processor_id();
 	kgdb_info[processor].debuggerinfo = regs;
-	kgdb_info[processor].task = current;
+	kgdb_info[processor].task = ipipe_safe_current();
 	atomic_set(procindebug[processor], 1);
 	atomic_set(kgdb_sync_softlockup[smp_processor_id()], 1);
 
@@ -884,8 +893,8 @@ int kgdb_deactivate_sw_breakpoints(void)
 	kgdb_break[i].saved_instr)))
 			return error;
 
-		if (CACHE_FLUSH_IS_SAFE  current-mm 
-addr  TASK_SIZE)
+		if (CACHE_FLUSH_IS_SAFE  ipipe_safe_current() == current 
+current-mm  addr  TASK_SIZE)
 			flush_cache_range(current-mm-mmap_cache,
 	addr, addr + BREAK_INSTR_SIZE);
 		else if (CACHE_FLUSH_IS_SAFE)
@@ -1069,7 +1078,7 @@ int kgdb_handle_exception(int ex_vector,
 		goto kgdb_restore;
 
 	kgdb_info[processor].debuggerinfo = linux_regs;
-	kgdb_info[processor].task = current;
+	kgdb_info[processor].task = ipipe_safe_current();
 
 	kgdb_disable_hw_debug(linux_regs);
 
@@ -1121,7 +1130,8 @@ int kgdb_handle_exception(int ex_vector,
 		*ptr++ = hexchars[(signo  4) % 16];
 		*ptr++ = hexchars[signo % 16];
 		ptr += strlen(strcpy(ptr, thread:));
-		int_to_threadref(thref, shadow_pid(current-pid));
+		int_to_threadref(thref,
+		 shadow_pid(ipipe_safe_current()-pid));
 		ptr = pack_threadid(ptr, thref);
 		*ptr++ = ';';
 
@@ -1213,7 +1223,8 @@ int kgdb_handle_exception(int ex_vector,
 			kgdb_hex2mem(remcom_in_buffer[1], (char *)gdb_regs,
  NUMREGBYTES);
 
-			if (kgdb_usethread  kgdb_usethread != current)
+			if (kgdb_usethread 
+			kgdb_usethread != ipipe_safe_current())
 error_packet(remcom_out_buffer, -EINVAL);
 			else {
 gdb_regs_to_regs(gdb_regs, linux_regs);
@@ -1334,7 +1345,8 @@ int kgdb_handle_exception(int ex_vector,
 /* Current thread id */
 strcpy(remcom_out_buffer, QC);
 
-threadid = shadow_pid(current-pid);
+threadid =
+	shadow_pid(ipipe_safe_current()-pid);
 
 int_to_threadref(thref, threadid);
 pack_threadid(remcom_out_buffer + 2, thref);
@@ -1488,7 +1500,8 @@ int kgdb_handle_exception(int ex_vector,
 			break;
 		case 'c':
 		case 's':
-			if (kgdb_contthread